MLflow — AI Glossary

Reference details

Related terms

• Experiment Tracking
• Model Registry
• Weights & Biases (W&B)

How it works

1. Runs are the unit: parameters, metrics and artefacts are logged against a run id, so an experiment is queryable rather than remembered.
2. Runs group into experiments, which is what makes two configurations comparable on recorded numbers instead of recollection.
3. Models are saved in a standard layout with a flavour declaring their framework, so a serving layer can load one without bespoke code.
4. A registry holds versioned models with stage transitions, which is where the approval lifecycle lives rather than in a distribution hub.
5. Tracing support records nested spans for model-backed applications, extending the same run-centred model to inference rather than only training.

Example: Why the run id is the whole discipline

Twelve experiments are run over a week with different parameters. Without logging, reproducing the best one means reconstructing which parameters were set when, from memory and shell history — and usually failing. With a run id, the parameters, metrics and artefact are a lookup. The cost is one logging call per run; the alternative cost is that the best result is a number nobody can produce again, which makes it worthless as a basis for the next decision.

Commonly confused with

Sits in the same sentence as model hubs and is a different thing. A hub distributes artefacts to consumers, optimising for sharing and discovery; a registry governs an internal lifecycle, optimising for who approved which version for which stage. This provides the registry and the experiment tracking around it. Against LLM-specific observability platforms the difference is lineage: this is organised around training runs and model versions, where they are organised around requests and traces, and a team doing both usually runs both.

When to use it

Reach for it when:

• Experiments must be reproducible from recorded parameters and metrics.
• An internal approval lifecycle is needed, with versioned stage transitions.
• Models should be packaged in a standard format a serving layer can load generically.

Reach for something else when:

• Only hosted APIs are called, where there is no training run or artefact to track.
• Per-request tracing of a model-backed application is the primary need.
• Distribution to external consumers, which is a hub's purpose.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to MLflow — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path